Crowd disasters have taken many lives. The Love Parade disaster (Duisburg, 2010), the Ellis Park Stadium disaster (Johannesburg, 2001), the PhilSports Stadium stampede (Manila, 2006) are just a few recent examples. Nevertheless, controlling crowds remains an unsolved problem, one that arises from the fact that in dense crowds, a “normal” situation may suddenly turn into a dangerous situation in the event of panic and that these changes are very difficult to predict as well as prevent.
The project “Detecting Anomalous Behavior in Stadium Crowds” was started in 2016. This project, led by Professor Sander Klous from the University of Amsterdam (UvA) and supported by the Netherlands eScience Center, attempts to model crowd behavior based on data gathered from the Wi-Fi signals of the smart phones of the participants in a crowd.
Dr Sonja Georgievska, eScience Research Engineer at the eScience Center, has been closely involved in the project, which recently ended. In the following interview, she talks about the aims and challenges of the project and looks back at the progress that was made.
Would you briefly describe the project and its aims?
The project was a collaboration between the University of Amsterdam, The Amsterdam ArenA (presently Johan Cruijff ArenA) Research and Innovation Lab, and the Netherlands eScience Center. It was one of the alliance projects, in which both the eScience Center and the project partners collaborate and mutually benefit from knowledge exchange, in this case regarding Big Data Analytics.
The scientific goal was to enable the detection of dangerous crowd behavior such as high-density spots, in real time, for the purposes of avoiding crowd disasters. Amsterdam ArenA was a so-called living lab, where data about the movements of concert visitors was anonymously collected via the Wi-Fi probing signals from their smart phones and the Wi-Fi network. The data collection and user localization analytics were performed by the team of Sander Klous, professor of Big Data Ecosystems for Business and Society.
What was your role in the project?
First, we wrote the project proposal together with the project partners. I subsequently worked on analytical methods for detecting dangerous crowd behavior based on the collected localization data.
How long were you involved?
About 1.5 years.
What were some of the challenges you faced in the project?
Our initial plans were to use data from crowd simulators to train machine learning algorithms to recognize dangerous situations. The trained neural network would then be able to recognize such situations live, when applied on the visitors’ locations as estimated from phone data. However, after analyzing the “localization” data, we quickly realized that our goals are unattainable due to the localization ambiguity observed from the data.
Digging deeper into this problem, I found out that it was in theory unsolvable. So we needed another plan. I subsequently proposed we use a probabilistic model instead, so as to be able to estimate the crowd density based on the data at hand. This approach, inspired by statistical mechanics, takes advantage of the large amount of data and bypasses the ambiguity problem. We also applied the same “big data” approach to other issues related to imperfections of the data source. However, this was not the end of the challenges. Another problem was to find ways to analyze experimentally our methodology, because a theoretical validation by itself cannot yield practical error estimates. Thus, most of my efforts and time, as well as that of the PhD student Philip Rutten, were spent on this last part.
Were the project aims achieved?
The solutions to the aforementioned problems were documented in the article “ Detecting high indoor crowd density with Wi‑Fi localization: a statistical mechanics approach ”, published this year in the Journal of Big Data. So yes, we are happy with the output.
What was the single most important insight you gained from this project (e.g. on crowd behavior and algorithmic prediction?
Despite the imperfections in the data source, in this case signals from smart phones, and despite some remaining theoretical limitations for tracking individual users, the availability of so much data, technology and the big-data-analytics approach allowed us to analyze the crowd as a whole. To do this, we combined classical mathematics from the 19th and 20th century, physics from the 20th century, and big data technologies from the 21st century in one package, which was very exciting in itself!
Read more about the project
Read more about Sonja Georgievska
